Skip to content

Latest commit

 

History

History
58 lines (42 loc) · 2.42 KB

README.md

File metadata and controls

58 lines (42 loc) · 2.42 KB

Multi-view ATPBind & ResiBoost

The official code implementation for Multi-view ATPBind model and ResiBoost algorithm from our paper, "Residue-Level Multi-View Deep Learning for Accurate ATP Binding Site Prediction and Its Applications in Kinase Drug Binding".

Here, we provide codes for training Multi-view ATPBind model with ResiBoost on ATPBind dataset and our processed kinase drug datasets.

Model description

The full model architecture and learning algorithm is shown below. model1

Multi-view ATPBind: The Multi-view model inputs both protein sequence and the corresponding 3D structure processed from pdb.

ResiBoost: The residue-level boosting algorithm performs boosting by undersampling poorly-predicted negative residues.

Model Training

You can train Multi-view ATPBind on ATPBind dataset from scratch using the command below.

$ python atpbind_main.py --model_keys esm-33-gearnet-resiboost --valid_folds 0

The resulting performances are written in the result file result_cv/result_cv.csv.

We also support training on 5 validation folds and different versions of models listed blow. To train on multiple validation sets or multiple versions, input the desired settings with separated with space;

$ python atpbind_main.py --model_keys esm-t33 esm-33-gearnet-resiboost --valid_folds 0 1 2 3 4

Supported training models

  • esm-t33 : ESM2 model (t33 version: 33layers, 650M params)
  • bert : ProtBERT model
  • gearnet: GearNet model
  • bert-gearnet: ProtBert+GearNet Multi-view model
  • esm-33-gearnet: ESM2+GearNet Multi-view model
  • esm-t33-ensemble: ESM2 + Mean Ensemble
  • esm-t33-resiboost: ESM2 + ResiBoost
  • bert-gearnet-ensemble: ProtBert+GearNet Multi-view model + Mean Ensemble
  • esm-33-gearnet-ensemble: ESM2+GearNet Multi-view model + Mean Ensemble
  • esm-33-gearnet-ensemble-rus: ESM2+GearNet Multi-view model + Random Undersampling
  • esm-33-gearnet-resiboost: ESM2+GearNet Multi-view model + ResiBoost

System Requirements

Hardware requirements

Multi-view ATPBind was trained using a server with 40 Intel(R) Xeon(R) Silver 4210R @ 2.40GHz CPUs, 128GB RAM and GeForce RTX 3090 GPUs. GPU memory usage was ~20GB.

Software requirements

Prerequisites

Multi-view ATPBind training and evaluation were tested for the following python packages and versions.

  • pytorch
  • torchdrug
  • numpy
  • pandas